A surrogate model to predict production performance in digital twin-based smart manufacturing

• Main Authors: Chua, Ping Chong, Moon, Seung Ki, Ng, Yen Ting, Ng, Huey Yuen
• Other Authors: School of Mechanical and Aerospace Engineering
• Format: Article
• Language: English
• Published: 2022
• Subjects: Engineering::Mechanical engineering; Digital Twin; Production Performance
• Online Access: https://hdl.handle.net/10356/162248
• Institution: Nanyang Technological University

With the dynamic arrival of production orders and unforeseen changes in shop-floor conditions within a production system, production scheduling presents a challenge for manufacturing firms to ensure production demands are met with high productivity and low operating cost. Before a production schedule is generated to process the incoming production orders, production planning is performed. Given the large number of input parameters involved in the production planning, it poses the challenge on how to systematically and accurately predict and evaluate production performance. Hence, it is important to understand the interactions of the input parameters between the production planning and the scheduling. This is to ensure that the production planning and the scheduling are coordinated and can be performed to achieve optimal production performance such as minimizing cost effectively and efficiently. Digital twin presents an opportunity to mirror the real-time production status and analyze the input parameters affecting the production performance in smart manufacturing. In this paper, we propose an approach to develop a surrogate model to predict the production performance using input parameters from a production plan using the capabilities of real-time synchronization of production data in digital twin. Multivariate adaptive regression spline (MARS) is applied to construct a surrogate model based on three categories of input parameters, i.e., current production system load, machine-based and product-based parameters. An industrial case study involving a wafer fabrication production is used to develop the surrogate model based on a random sampling of varying numbers of training data set. The proposed MARS model shows a high correlation coefficient and a large reduction in the number of input parameters for both linear and nonlinear cases with relation to three performances, namely flowtime, tardiness, and machine utilization.